“So Which Prompt Are We Actually Measuring With?”
Someone suggests checking whether the company turns up in AI, and the room agrees on the spot. Then you have to decide what to type in.
That is where we stalled, every time. Is it “best project management tools” or “project management tools compared”, and nobody could say why one over the other.
I got as far as a half-finished list of prompts and left it alone for two weeks. Plenty of companies have that same half-finished list open in a tab somewhere.
But the argument had a bad premise sitting underneath it.
The model doesn’t go and search the words you handed it. Somebody lined up 10,000 prompts against the search terms the assistants issued behind them, so that comparison is where this starts.
Does The Model Search The Words You Typed?
No. ChatGPT’s internally issued search terms shared 13% of their words with the original prompt, and when the same prompt went in a second time, 91% of the time a different search came out (Profound, 10,000 prompts over 14 days, published 30 April 2026) R.
Google describes the same mechanism in its own product: AI Mode splits a question into subtopics and fires off many searches at once, a design it calls query fan-out R.
Fan-out just means one question spreading into several searches.
Although Google explains the mechanism, it publishes no counts. Ordinary AI Mode gets “many”; Deep Search gets “hundreds”.
The figures below all come from the Profound side, which measured three engines and put numbers on each.
Three Engines, Three Different Amounts Of Your Wording Left
Perplexity stays close to what you asked: 88% word overlap, and only 14% of repeat runs produced a search it hadn’t already used.
Copilot sits in the middle at 50% and 47%. Line those up beside ChatGPT’s 13% and 91% and “we measure it in AI” stops being one activity.
Where the three do agree is volume. Every one of them issued roughly 1.4 to 2 searches per run.
Same number of notes taken, wildly different handwriting. One person copies you down word for word; the next paraphrases you afresh every single time.
What Survives The Rewrite: Your Name And Your City
The same study sorted the parts of a prompt by how well they held up.
- Brand names come through largely intact on all three engines
- Location is preserved almost without exception, on all three
- Price bands and other filters get reinterpreted or widened by ChatGPT, and are often dropped outright by Perplexity and Copilot
That first line is good news for the measurement most people want. Ask an engine about you by name, and the name is the part that reaches the search.
For example, take “under $50” or “for beginners”. Since those qualifiers are the first thing to fall off, the answer the reader sees may have been assembled from results where the qualifier was already gone — so you can be the strongest option under $50 and have that fact never make it into the sentence.
Six Lines To Put On The Slide Before The Number
The engines covered are ChatGPT, Perplexity and Copilot. Gemini and Google AI Mode are outside it. The Google explanation quoted earlier backs up the mechanism; it isn’t where these percentages came from.
Second, the makeup of the 10,000 prompts stays closed — the language mix, the industries, the intents. Since you can’t see whether prompts from your own market are in the pile, carrying the percentages straight into your own category is early.
The observation window is 14 days, which is short. The method for capturing fan-out queries isn’t disclosed either.
And Profound sells AI-visibility measurement. “The prompts you assumed aren’t enough” is a conclusion that points at what they sell, which doesn’t make it wrong, though it belongs on the same slide as the number.
Values like “1.4 to 2 searches” will move the day an engine changes its implementation.
I started writing those six lines into the notes column before pasting any of the figures in.
Ask The Same Thing Several Ways, Then Count How Often Your Name Comes Back
Two things change in how you measure, and neither of them is large.
The first is to build several prompts around your brand name and your location, ask them repeatedly, and read the result as a rate of appearance. Since ChatGPT sends out a different search 91% of the time, one prompt asked once can’t settle whether you show up.
The second is to measure your qualified strengths in their own bucket. Assume the price band and the audience filter will fall off, keep those prompts separate from the unqualified ones, and you’ll still be able to read the result correctly a month later.
If you’re stuck on which prompt to open the tracking with, the “best X” shape is the easiest doorway. It holds its form better than most: the original question survived intact in 65% of Perplexity runs, 52% of Copilot’s and 39% of ChatGPT’s.
How anyone estimates the prompts real people type is its own problem, and so is what a “prompt search volume” figure is actually worth. Whether the phrasing changes the answer you get is over here.
Today’s question sits one floor below all three: what the model does to the prompt you picked, before it searches at all. Line up three prompts and start counting.
